diff --git a/config/ucx.conf b/config/ucx.conf
index d40117752a3..98813d65bab 100644
--- a/config/ucx.conf
+++ b/config/ucx.conf
@@ -6,6 +6,7 @@ UCX_IB_MLX5_DEVX_OBJECTS=
 UCX_GDR_COPY_BW=0MBs,get_dedicated:30GBs,put_dedicated:30GBs
 UCX_GDR_COPY_LAT=30e-9
 UCX_DISTANCE_BW=auto,sys:16500MBs
+UCX_CUDA_COPY_BW=h2d:400GBs,d2h:300GBs,d2d:400GBs,other:10000MBs
 
 [Fujitsu ARM]
 CPU vendor=Fujitsu ARM
diff --git a/src/uct/cuda/cuda_copy/cuda_copy_iface.c b/src/uct/cuda/cuda_copy/cuda_copy_iface.c
index 49b45df9bf1..da5f4fb303c 100644
--- a/src/uct/cuda/cuda_copy/cuda_copy_iface.c
+++ b/src/uct/cuda/cuda_copy/cuda_copy_iface.c
@@ -37,14 +37,24 @@ static ucs_config_field_t uct_cuda_copy_iface_config_table[] = {
      "Max number of cuda events. -1 is infinite",
      ucs_offsetof(uct_cuda_copy_iface_config_t, max_cuda_events), UCS_CONFIG_TYPE_UINT},
 
-    {"BW", "10000MBs",
-     "Effective memory bandwidth",
-     ucs_offsetof(uct_cuda_copy_iface_config_t, bandwidth), UCS_CONFIG_TYPE_BW},
+    /* TODO: 1. Add separate keys for shared and dedicated bandwidth
+             2. Remove the "other" key (use pref_loc for managed memory) */
+    {"BW", "10000MBs,h2d:8300MBs,d2h:11660MBs,d2d:320GBs",
+     "Effective memory bandwidth", 0,
+     UCS_CONFIG_TYPE_KEY_VALUE(UCS_CONFIG_TYPE_BW,
+         {"h2d", "host to device bandwidth",
+          ucs_offsetof(uct_cuda_copy_iface_config_t, bw.h2d)},
+         {"d2h", "device to host bandwidth",
+          ucs_offsetof(uct_cuda_copy_iface_config_t, bw.d2h)},
+         {"d2d", "device to device bandwidth",
+          ucs_offsetof(uct_cuda_copy_iface_config_t, bw.d2d)},
+         {"other", "any other memory types combinations bandwidth",
+          ucs_offsetof(uct_cuda_copy_iface_config_t, bw.other)},
+         {NULL})},
 
     {NULL}
 };
 
-
 /* Forward declaration for the delete function */
 static void UCS_CLASS_DELETE_FUNC_NAME(uct_cuda_copy_iface_t)(uct_iface_t*);
 
@@ -134,7 +144,7 @@ static ucs_status_t uct_cuda_copy_iface_query(uct_iface_h tl_iface,
 
     iface_attr->latency                 = UCT_CUDA_COPY_IFACE_LATENCY;
     iface_attr->bandwidth.dedicated     = 0;
-    iface_attr->bandwidth.shared        = iface->config.bandwidth;
+    iface_attr->bandwidth.shared        = iface->config.bw.other;
     iface_attr->overhead                = UCT_CUDA_COPY_IFACE_OVERHEAD;
     iface_attr->priority                = 0;
 
@@ -407,16 +417,17 @@ uct_cuda_copy_estimate_perf(uct_iface_h tl_iface, uct_perf_attr_t *perf_attr)
         perf_attr->bandwidth.dedicated = 0;
         if ((src_mem_type == UCS_MEMORY_TYPE_HOST) &&
             (dst_mem_type == UCS_MEMORY_TYPE_CUDA)) {
-            perf_attr->bandwidth.shared = (zcopy ? 8300.0 : 7900.0) * UCS_MBYTE;
+            perf_attr->bandwidth.shared = zcopy ? iface->config.bw.h2d :
+                                          iface->config.bw.h2d * 0.95;
         } else if ((src_mem_type == UCS_MEMORY_TYPE_CUDA) &&
                    (dst_mem_type == UCS_MEMORY_TYPE_HOST)) {
-            perf_attr->bandwidth.shared = (zcopy ? 11660.0 : 9320.0) *
-                                          UCS_MBYTE;
+            perf_attr->bandwidth.shared = zcopy ? iface->config.bw.d2h :
+                                          iface->config.bw.d2h * 0.95;
         } else if ((src_mem_type == UCS_MEMORY_TYPE_CUDA) &&
                    (dst_mem_type == UCS_MEMORY_TYPE_CUDA)) {
-            perf_attr->bandwidth.shared = 320.0 * UCS_GBYTE;
+            perf_attr->bandwidth.shared = iface->config.bw.d2d;
         } else {
-            perf_attr->bandwidth.shared = iface->config.bandwidth;
+            perf_attr->bandwidth.shared = iface->config.bw.other;
         }
     }
 
@@ -491,7 +502,10 @@ static UCS_CLASS_INIT_FUNC(uct_cuda_copy_iface_t, uct_md_h md, uct_worker_h work
     self->id                     = ucs_generate_uuid((uintptr_t)self);
     self->config.max_poll        = config->max_poll;
     self->config.max_cuda_events = config->max_cuda_events;
-    self->config.bandwidth       = config->bandwidth;
+    self->config.bw.h2d          = config->bw.h2d;
+    self->config.bw.d2h          = config->bw.d2h;
+    self->config.bw.d2d          = config->bw.d2d;
+    self->config.bw.other        = config->bw.other;
     UCS_STATIC_BITMAP_RESET_ALL(&self->streams_to_sync);
 
     ucs_mpool_params_reset(&mp_params);
diff --git a/src/uct/cuda/cuda_copy/cuda_copy_iface.h b/src/uct/cuda/cuda_copy/cuda_copy_iface.h
index 700345d22a2..1b54298dfaa 100644
--- a/src/uct/cuda/cuda_copy/cuda_copy_iface.h
+++ b/src/uct/cuda/cuda_copy/cuda_copy_iface.h
@@ -69,7 +69,12 @@ typedef struct uct_cuda_copy_iface {
     struct {
         unsigned                max_poll;
         unsigned                max_cuda_events;
-        double                  bandwidth;
+        struct {
+            double              h2d;
+            double              d2h;
+            double              d2d;
+            double              other;
+        } bw;
     } config;
     /* handler to support arm/wakeup feature */
     struct {
@@ -87,7 +92,12 @@ typedef struct uct_cuda_copy_iface_config {
     uct_iface_config_t      super;
     unsigned                max_poll;
     unsigned                max_cuda_events;
-    double                  bandwidth;
+    struct {
+        double              h2d;
+        double              d2h;
+        double              d2d;
+        double              other;
+    } bw;
 } uct_cuda_copy_iface_config_t;